Can our payment institutions innovate?

Banks earning season is rolling in with Q2/2017 earnings. Most major banks including Bank of America, Wells Fargo, Goldman Sachs etc have lowered their guidance for the rest of the year citing lots of reasons - trading, market making, interest rates etc. Mortgage and lending business is also under pressure it seems. The same time Netflix reported record consumer growth in Q2/2017 while also suggesting strong negative free cash flows for the suggestible future. The market was happy at that and sent the stock surging. Its just part of doing business, right? That just sounds too bad for U.S banks which are considered fundamental pillars of the economy. Fundamental enough that taxpayers bailed them out few years back but still not allowed to innovate.

Bitter truth is that innovation in core parts of the economy can make the economy unstable. Innovation implies risk and risk can materialize once in a while. So we allow innovation to happen on the fringes of the economy. Fringe enough that if we loose it, not many people would notice.

Back to FinTech or Innovation in financial services. In the U.S/EU FinTech is helping evolve micro-lending and online banking industries. The pace is much more rapid in the developing parts of the world though, like Africa. However, we may be surprised to hear that U.S is a laggard when it comes to financial innovation. The reason is not that innovators in the U.S are not smart enough - but they simply focus on other important problems like talking robots, space travel and auto driving cars.

So what is going on with FinTech that should worry us? Here is a news flash - China is fast becoming the new FinTech capital of the world, taking that title away from the western world. The transactions volumes processed by Baidu, Tencent and Alibaba are growing so fast that they are set to pass the volumes Visa and Mastercard process in Europe and US in 2017. China's mobile payment system is does not use any high-technology though. It uses bar codes to process payments! Thats way inferior and insecure than the system used by Mastercard and Visa (EMVCO) or by Google/Applepay. One thing it gets right though - its fast, convenient and cheap and thats what matters. It has connected about a billion people in China who never had any bank accounts or credit history before. So why should we be worried? Chinese companies are slowly importing this system into U.S and EU through partnerships and deals with processors like Stripe and companies like AirBnB etc. They are so big and powerful that U.S companies have little choice accommodating them. That would slowly cut into the western financial system and would cause a huge disruption in the economies in the next ten years.

What can we learn from the success of the Chinese system? First, it processes transactions without communicating with the banks each time. If the amount of the transaction has a limit and the replication is fast enough, this should work. The result is that this makes it very fast for high volume processing.  It is cheap and convenient because there is no need to have expensive NFC hardware or scanning devices or chips/cards. Compare that with the our system - there are two banks involved. We use a dedicated transmission network for the payments and there are middlemen involved. More middlemen mean more fees and commissions including the need to keep reconciling data as it passes all the intermediate systems. Chip cards make things even slower. Evolution in our system is not going to be easy since the credit card companies have a tight grip on end users with rewards and points systems.

What can save us from certain failure is the need for innovation - the U.S still has around 50 million people without bank accounts and thus no credit cards. PayNearMe tapped into this market by offering an easy way to make cash based payments where credit cards were not an option. So who is going to get to it first!

How social networking combined with NLP analytics is helping expand the economy

Recently I have been researching the rise of NLP again. This was the topic of my bachelor thesis in 1995, almost 20 years ago and it has now become a hot area of research again in the last 5 years. The science and tools have evolved and a lot of new open source tools like NLTK are available for researchers.

Clearly the early users of social networking data were doing a lot of sentiment analysis on it to determine trends for companies, products, politics etc. Things have changed now - governments are interested in scouring billions of bytes of data generated daily in social networks for intelligence hints, Fin-Tech upstarts are starting to successfully use the same data to disrupt financial services - for example, the lending industry. 

The sale of Troo.ly to Airbnb spiked my interest into this. Research revealed the existence of a whole bunch of companies and existing patents in the science of using social network data to determine a person's trust or risk score. Appears that lot of tricks have evolved in the last five years. However, this is quite scary as well! In the next few years, I can see a lot of people trying to use this score instead of just pulling the credit scores in business transactions. This could be landlords, people selling goods on craigslist (e.g cars etc), small business engaging in seasonal hiring to keep costs low. It could certainly replace simple background checks. While this is all great, the scary part is that this presents a Kafkian situation for people at the other end of it. Unlike their FICO scores, there is a lot less visibility into how the machine learning algorithms work. Imagine a customer service rep trying to explain why your social risk was considered too high when renting an apartment for a few days on AirBnB.

This led me to findings how the big Chinese payment companies (Ant Financial, Baidu) are now planning to use all the data that they collect from repeated payment transactions and the social networks they own to determine credit scores of people. This idea is definitely not new but it is breaking new ground in the last couple of years. For example, companies like HelloSoda, Guardian Analytics have been doing risk scoring for quite some time now. There are also a lot of banking and lending upstarts in U.S and EU - Kabbage, Simple, Moven, Fidor etc. However, FinTech has taken social networking data to a new level now and this time its not just to send marketing and sales offers (like the famous American Express examples show) or to make your banking very cool indeed. 

There are multiple innovation areas - companies like Jumo have made a significant penetration in the underbanked markets of Africa to enable micro-lending based on social networking data that they collect. In China, Baidu already has more than 10% of its assets involved in some kind of lending, the biggest being in the education market. Now Alibaba and Tencent are also getting involved. Not only is the mobile payments market in China set to pass the transactions Visa and Mastercard process but it is also spinning out new uses of the data collected. In one way it looks like social networking companies may have an edge over managing risk and may have slowly started to disintermediate traditional lending companies. There are 800 million people using Tencent for payments that have no credit history with the central bank and the daily transaction patterns reflect a lot about their behavior and risk. Take a look at AutoGravity - users can not only select cars, schedule a test drive but once they connect it to their social networking account, the site also prefills the application, verifies identity and lines up four lending companies without having the user fill up scores of forms or go to an office. All on the mobile phone in minutes. Facebook, Google and Apple could be doing this next year. It all starts by focusing on an underserved market and there are enough in financial services - there are almost 64M Americans without sufficient credit history and almost 2B people around the world without a bank account.

Learning TensorFlow

TensorFlow is Google's deep learning library, released in 2015. In many ways, it may be puzzling why we should pay attention to it given there are so many machine learning frameworks around that seem to be doing a pretty good job so far!

The following should provide a good motivation:

- TensorFlow supports GPUs.

- TensorFlow supports distributed computation

- Primarily TensorFlow is good for deep learning. Lets just say it seems to be much more focused on DL.

Note that TensorFlow is equivalent to the numpy module in python. There is a lot of development still going on and hopefully easy to use libraries like scikit-learn will be available soon. One may also ask that Apache Spark provides distributed computation, has an ML library and supports GPUs as well. So why not just use Spark by itself? The answer may be that Spark is not focused on DL as much as TF is. Moreover the distributed computation model of Spark is very different from TensorFlow. Spark has a resource manager hidden from the user that parallelizes an RDD computation over a cluster. TensorFlow distributed programming involves the user and the program has a lot more control on the computation. IMO, Spark may sit in the data pipeline ahead of TensorFlow to massage/clean and process data that is used to train a very large neural network. At this point, TensorFlow needs a considerable simplification of its cluster management and programming API before it can be used by data scientists used to working with tools like numpy/R or Spark.

Here are some good talks and links to understand TensorFlow better:

• TensorFlow Basics and library comparisons
• Distributed TensorFlow
• TensorFlow architecture
• Good compilation of TensorFlow tutorials, books and other links

Why Logistic Regression is linear even though it uses a non linear function

Logistic Regression uses a non linear activation function - the logistic function:

\[z = \frac{1}{(1 + e^{-y})}\]

where $\textit{y}$ is linear in $\textit{x}$, the input variable.

Note that this is equivalent to:

\[z = \frac{e^{y}}{(1 + e^{y})}\]

So why is logistic regression considered linear and the result used for classification rather than predicting a continuous output? Having more of a computer science background, this was something that did not initially catch my eye. This related post made it quite easy for me to understand logistic regression. Here I provide some key points related to logistic regression and some references from a theoretical perspective to help develop a better understanding:

• Logistic function is used to give the probability of the output being in a binary class. Its output is always between 0 and 1 given any value of inputs in any number of dimensions.
• If you rearrange the logistic function, the natural log of the odds (the ratios of probabilities of an event being successful and unsuccessful) is the familiar linear regression equation. The reason why logistic regression is considered linear is that we are combining the outputs using a linear function.
• The tanh function, which is a mathematical function of the logistic function is a better choice than logistic function since it has steeper gradient. The steeper gradient is better in backprop training. A steeper gradient passes back feedback from output back to input much faster and having a larger impact on weights closer to input nodes making convergence faster.
• While logistic regression works very well in binary classification for any number of dimensions, the softmax function is a much better choice for in multi-class classification. The softmax sums up to one in a multi-class situation over all the classes. The logistic function does not have this property. 
• In a binary classification, using a softmax function is equivalent to using the sigmoid function.
• One may ask why use these complicated exponential functions in the output? If we want probability, we can simply use an average - divide each output by the sum of outputs. The problem with this approach is that individual values can become negative even if they add to one. Exponentiating makes everything positive. Also, exponentiation works well for back propagation since it amplifies errors making algorithms converge faster.
• The use of logistic function as an activation function inside the network also has an issue - the vanishing gradient problem which makes deep neural nets very hard to train. ReLU, y = max(x, 0), has been a popular choice since it does not alter the gradients as they are propagated back to the input. This nice blog entry provides a great explanation.
• ReLU is also only used for hidden layers. Outputs would still be softmax (classification) or linear (regression).

Better ways to do One Hot Encoding

While running an ML algorithm on any data, you may have to convert categorical data into numerical data - reason is that a mostly all scikit-learn code requires you to input data which is numeric. Though one may think that it is a scikit limitation, that may not be true. Since ML uses math and vectors behind the scenes, the data has to be numerical for most good algorithms.

One of the common ways to convert categorical data to numeric data is using One Hot Encoding. This kind of encoding uses indicator variables, where each value of the category is replaced by a column of its own. This can lead to column explosion so one must be careful. A lot of times, categories that have an order can be mapped to numerical values that may be helpful as well. 

Several methods of One Hot Encoding have been mentioned. Most prominent and simple of them uses the get_dummies function in pandas:

In this function, we loop through all the categorical variables in the pandas dataframe one by one and for each case, we use get_dummies to create indicator variable columns (which are numeric) and then we delete the original categorical column. This is pretty simple and mostly this is what is recommended the most in forums. However, there are a few catches with this method in practice:

1. The categorical variable may have different sets of values in training, validation and test data. If you run get_dummies, it may assign the same numeric value to different categorical values. When fed into the ML algorithm this can cause unintended data alternation and results.
2. The order in which the categorical values are encountered by get_dummies may further contribute to the above issue
3. Some categorical values may only appear in validation and test data and may be absent in training data. This can cause more problems. Training a model on one kind of values and then testing predictions on another may not make sense. The get_dummies does not help identify this problem. 

The best course of action is to map the categorical values to a set of indicator variables and then use the same set during test and validation. This should not be changed. In addition, if certain categorical values that are going to be seen in practice in validation or test data, we must take that into account. We will leave this specific problem to another post. In this post, lets see how we can fix get_dummies to at least fix the first two problems and alert us of the third one.

Python provides few other alternatives that are a bit complex to use but I feel are totally necessary. One of these uses the LabelEncoder function. LabelEncoder looks at a categorical variable and creates a transformation which maps the values to integer labels. This does not create any indicator variables so one may think is inadequate for our needs. However, what LabelEncoder does is store the mapping as a model which can be used repeatedly later on. Combining LabelEncoder with get_dummies provides the ideal solution:

So what's up with the dictionary and the train variable in this function? Note our initial objective - we must use the same mapping for both the training and test data. The python dictionary holds important storage area for the mappings created by the LabelEncoder for each column in a dataframe. The call made to this function for training data looks as follows:

train_data, le_dict = oneHotEncode2(train_data)

Then on the test data, the call is made by passing the dictionary returned back from training:

test_data, _ = oneHotEncode2(test_data, le_dict)

The call to an already created transform for encoding also checks to see if it encounters any new values in the test data. If it does, it will warn us and we can go back and take appropriate action.

Hail Seaborn!

The seaborn heatmap perhaps is the best visualization of the correlations in a data set.

 Much better than Axes.matshow():

Analyzing Predictions to find *bugs*

Testing should be one of the most frequent steps in the SDLC for any code, whether it is building a website, powering a smart device or machine learning. Tests of machine learning code may reveal lots of issues - transformations that don't work because the right columns are not operated on or data which fails assumptions of the learning algorithm. The list may be endless and we are well attuned to this process. One of the tests often ignored is comparing the test results during the training period with actual values from the data and then trying to find what went wrong and where. This can often reveal attributes or column values that you may have ignored till now and must play a bigger role.

For example, use the following function to plot test results against the real values. The plot should be a straight line. Any deviations from it are to be analyzed for every column.

Make your test data comply with training data

When using scikit-learn libraries with pandas, you would often get errors if the test data does not have columns which match the model created from the training data set. You may have deleted columns that you did not need from the training data when creating the model or you may have constructed new columns based on existing data variables (for example, creating 'Age' from 'Date' or combining the effects of multiple variables). In most cases, this is simple to achieve by wringing the test data through the same function as the training data. However, when you are using One-Hot-Encoding of categorical data, the columns created as a result of this coding in the training set and the test set may not match for the simple reason that some values for categorical data may only be present in the training data set and others may be present only in the test data set. I have included here a python function that I wrote and works really well to match the test data with the training data:

K-Means Cluster Analysis on Outlook on Life Survey Data

The Outlook on Life Survey (OOL) was designed to study political and social attitudes in the United States. The project included two surveys fielded between August and December 2012 using a sample from an Internet panel. A total of 2,294 respondents participated in this study during Wave 1 and 1,601 were interviewed during Wave 2. There are 436 variables in this study.

Note that full access to data archive is available for members on the ICPSR site.

Our goal in this study is to do a K-Means Cluster Analysis to determine if there is a set of variables we can identify that can measure how angry people feel about the way things are going in the country these days (W1_B4).

The variables we selected as the cluster variables are:

Variable	Description
W1_H1	Society has reached the point where Blacks and Whites have equal opportunities for achievement
W1_O5	[Black people should teach their children to be careful around the police] How much emphasis or de-emphasis should Black people place on each statement in the education of their children?
PPINCIMP	Household Income
W1_F4_D	[To become wealthy ] For yourself and people like you, how easy or hard is it to reach these goals?
W1_F6	How far along the road to your American Dream do you think you will ultimately get on a 10-point scale where 1 is not far at all and 10 nearly there?
W1_M5	How often do you attend religious services?

We are going to validate our clusters by excluding the variable W1_B4 (Description - Generally speaking, how angry do you feel about the way things are going in the country these days?) from our analysis. We would treat this variable as a measure and we expect to see some differences on this variable amongst our clusters.

While doing the cluster analysis we noticed the following: (a) In some cases, when more data was included for determining the clusters or the cluster analysis was repeated, we observed that the cluster means for our measure variable came quite close. In addition, if the number of clusters we wanted was increased, that also had an impact on the cluster means and the significance of the difference amongst the cluster means.

First, we ran a simulation trying to determine what the affect of clustering on the average distance of the points from centroids. For our analysis, we will go with three clusters. As we can see the choice of number of clusters is completely subjective. As the number of clusters increase, the average distance will decrease.

Here is the plot of the cluster after applying PCA to reduce the number of variable dimensions to two:

The split amongst the three clusters is (the cluster IDs are 0.0, 1.0 and 2.0)

0.0    419

2.0    355

1.0    348

Our next step is to try and fit a simple least squares line using W1_B4 as the dependent variable based on the cluster number we have determined for the points. The results of that are:

The mean and standard deviation for W1_B4 in the three clusters are:

cluster        
0.0      2.916468
1.0      3.129310
2.0      3.008451
standard deviations for W1_B4 by cluster
            W1_B4
cluster        
0.0      1.307026
1.0      1.192171
2.0      1.224716

Because we had more than two clusters, we decided to conduct a tukey test to determine which one of the two means had significant difference. It seems that cluster 0.0 and cluster 1.0 had means that were significantly different (led to rejection of the null hypothesis).

Then we analyzed the two clusters are the means of the different cluster variables within those clusters.

The following is the mean of the various clustering variables in their respective clusters:

W1_B4 mean is lower for cluster 0.0 indicating that people in cluster 0.0 are generally more angry about the situation in the country than those in cluster 1.0 (lower values in the survey correspond to higher anger felt by the respondent). Then we try to look deeper into the cluster variable means. We find that W1_F4_D is lower for cluster 0.0. Lower value indicates that the respondents think that it is hard to become wealthy in this country. A lower value of W1_H1 (in cluster 0.0) indicates that the respondents agree much more that blacks and whites have equal opportunities for achievement. A lower value of W1_M5 for cluster 0.0 also indicates that respondents attend religious services more frequently there. A higher value for W1_05 for cluster 0.0 indicates that respondents place a strong emphasis that black people should teach their children to be careful around the police in the education of their children. Cluster 0.0 also has lower average household income.

So what can we establish - it seems that the likelihood that a person is going to be unhappy about the situation in this country is going to be higher when their income is lower and they generally do not believe that their wealth can significantly increase in this country (even though they believe that blacks and whites have equal opportunities). The higher likelihood seems connected with lower income, more religious bent of mind and general wariness of law enforcement.

The code is included below:

Lasso Regression for U.S. National Epidemiological Survey on Alcohol and Related Conditions (NESARC) data

Background on the data used here (NESARC) was described in an earlier post in this blog. In that post, we conducted a Random Forest analysis for the boolean response variable: EVER GAMBLED 5+ TIMES IN ANY ONE YEAR (S12Q1). In this post we conduct Lasso Regression for the same. Before we dive into the analysis, there were some interesting observations about using Lasso, which performs linear regression for a response variable. When this response variable has an integer value vs a boolean we get different results. When we use a boolean variable, the results are quite similar to the Random Forest analysis which also used a boolean response variable. However, if we use an integer value for the response variable, which makes more sense to use in a linear regression (e.g Lasso where MSE is calculated), the list of important variables detected are different from those reported by Random Forest. This requires more analysis.

The following variables (listed below in order of importance) were identified as the most important since their coefficients were the largest in absolute values in this regression. We can compare the coefficients since we scaled the input variables to zero mean and unit variance. The most important variables identified in this analysis were: whether the individual had panic attacks (S6Q1) or was engaged in reckless driving (S11AQ1A15) or was not too open with anyone including those close to him/her (S10Q1A3). The rest of the variables are mentioned as well. The variables whose coefficient was reported to be zero by Lasso are briefly mentioned below as well.

Variable	Description	Coefficients
S6Q1	HAD PANIC ATTACK, SUDDENLY FELT FRIGHTENED/OVERWHELMED/NERVOUS AS IF IN GREAT DANGER BUT WERE NOT	0.226163
S11AQ1A15	EVER GET MORE THAN 3 TICKETS FOR RECKLESS/CARELESS DRIVING, SPEEDING, OR CAUSING AN ACCIDENT	0.175727
S10Q1A3	FIND IT HARD TO BE "OPEN" EVEN WITH PEOPLE YOU ARE CLOSE TO	0.125577
S11AQ1A25	EVER DO SOMETHING YOU COULD HAVE BEEN ARRESTED FOR, REGARDLESS OF WHETHER YOU WERE CAUGHT OR NOT	0.085541
S11AQ1A2	EVER STAY OUT LATE AT NIGHT EVEN THOUGH PARENTS TOLD YOU TO STAY HOME	0.068325
S11AQ1A22	EVER SHOPLIFT	0.067515
S11AQ1A1	OFTEN CUT CLASS, NOT GO TO CLASS OR GO TO SCHOOL AND LEAVE WITHOUT PERMISSION	0.067146
S11AQ1A14	EVER DO THINGS THAT COULD EASILY HAVE HURT YOU OR SOMEONE ELSE, LIKE SPEEDING OR DRIVING AFTER HAVING TOO MUCH TO DRINK	0.054445
S9Q1A	EVER HAD 6+ MONTH PERIOD FELT TENSE/NERVOUS/WORRIED MOST OF TIME	0.047805
S10Q1A16	THE KIND OF PERSON WHO FOCUSES ON DETAILS/ORDER/ORGANIZATION OR LIKES TO MAKE LISTS AND SCHEDULES	0.047434
S1Q1G	NUMBER OF YEARS LIVED IN UNITED STATES	-0.030454
SMOKER	TOBACCO USE STATUS	0.026548
S11BQ1	BLOOD/NATURAL FATHER EVER HAD BEHAVIOR PROBLEMS	0.024917
AGE	CYEAR (DATE OF INTERVIEW: YEAR) - DOBY (DATE OF BIRTH: YEAR)	-0.024049
S2AQ5G	HOW OFTEN DRANK 5+ BEERS IN LAST 12 MONTHS	0.022640
S10Q1A25	HAVE OTHERS TOLD YOU THAT YOU ARE STUBBORN OR RIGID	0.019741
S3AQ3B2	USUAL FREQUENCY WHEN SMOKED CIGARS	-0.015306
S1Q1D5	"WHITE" CHECKED IN MULTIRACE CODE	-0.012658
S1Q10A	TOTAL PERSONAL INCOME IN LAST 12 MONTHS	-0.012306
MARITAL	CURRENT MARITAL STATUS	0.011049
DGSTATUS	DRUG USE STATUS	-0.009435
S3AQ3B1	USUAL FREQUENCY WHEN SMOKED CIGARETTES	0.008346
S1Q9B	OCCUPATION: CURRENT OR MOST RECENT JOB	-0.007877
CHLD0_17	NUMBER OF CHILDREN UNDER AGE 18 IN HOUSEHOLD	0.007562
S2AQ10	HOW OFTEN DRANK ENOUGH TO FEEL INTOXICATED IN LAST 12 MONTHS	-0.007253
S10Q1A43	ARE THERE VERY FEW PEOPLE YOU'RE REALLY CLOSE TO OUTSIDE OF IMMEDIATE FAMILY	0.007089

Variables reported with zero coefficents (these could be correlated with the variables mentioned above that have non-zero coefficents since Lasso would randomly end up picking one of the correlated variables) or unrelated with the response variable:

Variable	Description	Possible Explanation
S10Q1A52	THE SORT OF PERSON WHO DOESN'T CARE ABOUT WHAT PEOPLE THINK OF YOU	Seems Correlated
S10Q1A58	FLIRT A LOT	Seems Correlated
S2AQ12F	HOW OFTEN DROVE MOTOR VEHICLE AFTER 3+ DRINKS IN LAST 12 MONTHS	Seems Correlated
S10Q1A47	HAVE ALMOST ALWAYS PREFERRED TO DO THINGS ALONE RATHER THAN WITH OTHERS	Seems Correlated
S10Q1A46	TAKE LITTLE PLEASURE IN BEING WITH OTHERS	Seems Correlated
S10Q1A45	WOULD BE JUST HAPPY WITHOUT HAVING ANY CLOSE RELATIONSHIP	Seems Correlated
S2AQ12B	HOW OFTEN DRANK AFTER MIDNIGHT IN LAST 12 MONTHS	Seems Correlated
S10Q1A32	OFTEN GET ANGRY OR LASH OUT WHEN SOMEONE CRITICIZES OR INSULTS YOU	Seems Correlated
S10Q1A22	HARD TO LET OTHERS HELP IF THEY DON'T AGREE TO DO THINGS EXACTLY THE WAY YOU WANT	Seems Correlated
S2AQ9	HOW OFTEN DRANK 4+ DRINKS OF ANY ALCOHOL IN LAST 12 MONTHS (WOMEN ONLY)	Seems Correlated
NUMPER18	NUMBER OF PERSONS 18 YEARS AND OLDER IN HOUSEHOLD	Seems Unrrelated
NUMPERS	NUMBER OF PERSONS IN HOUSEHOLD	Seems Correlated/Unrelated

The training and test R-square errors were reported as 0.670503947 and 0.66454160107 respectively which were fairly high indicating a good fit of the model with the training and test data. The training and test mean squared error was 0.474361978873 and 0.495379372341 respectively which was fairly low in both cases as well. The MSE plot is as follows and we can see that the MSE is successively going down as the alpha value increases.

The plot for the progression of the coefficients as the variables are added one-by-one in the Lasso Regression is show below.

Python code is included below:

Random Forest Analysis of U.S. National Epidemiological Survey on Alcohol and Related Conditions (NESARC) data

In 2001/2002, the National Institute on Alcohol Abuse and Alcoholism (NIAAA) conducted the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC), the largest and most ambitious comorbidity (simultaneous presence of two chronic conditions in a patient) study ever conducted. In addition to an extensive battery of questions addressing present and past alcohol consumption, alcohol use disorders (AUDs), and utilization of alcohol treatment services, NESARC included similar sets of questions related to tobacco and illicit drug use (including nicotine dependence and drug use disorders).

The unprecedented sample size of NESARC (n = 43,093) made it possible to achieve stable estimates of even rare conditions. Moreover, its oversampling of Blacks and Hispanics as well as the inclusion of Hawaii and Alaska in its sampling frame yielded enough minority respondents to make NESARC an ideal vehicle for addressing the critical issue of race and/or ethnic disparities in comorbidity and access to health care services.

NESARC studies the occurrence of more than one psychological disorder or substance use disorder in the same person. In this analysis I have utilized the data collected in the survey for gambling and tried to determine its linkage to other predictors like alcoholic tendencies or drug use or other personality traits that can be detected. General information about the survey can be found here.

The following response variable was predicted: EVER GAMBLED 5+ TIMES IN ANY ONE YEAR (S12Q1).

The following variables (listed below in order of importance) were amongst those included as explanatory variables in Random Forest technique. The accuracy of the random forest was 74% with the maximum accuracy achievable with around 12 trees in the forest. Further increase in number of estimators (trees) did little to increase overall accuracy of the model. On closer examination of the variables found important by the model, we can see that an we can predict about the gambling tendencies of an individual by examining their alcohol consumption patterns, personality and drug use amongst other things.

Variable	Description	Score
S1Q1G	NUMBER OF YEARS LIVED IN UNITED STATES	0.059010
S1Q10A	TOTAL PERSONAL INCOME IN LAST 12 MONTHS	0.057991
AGE	CYEAR (DATE OF INTERVIEW: YEAR) - DOBY (DATE OF BIRTH: YEAR)	0.056678
S1Q1E	ORIGIN OR DESCENT	0.048400
S1Q9B	OCCUPATION: CURRENT OR MOST RECENT JOB	0.046075
BUILDTYP	TYPE OF BUILDING FOR HOUSEHOLD	0.036685
MARITAL	CURRENT MARITAL STATUS	0.029524
SMOKER	TOBACCO USE STATUS	0.028936
NUMPERS	NUMBER OF PERSONS IN HOUSEHOLD	0.027357
S2AQ12B	HOW OFTEN DRANK AFTER MIDNIGHT IN LAST 12 MONTHS	0.025185
S10Q1A52	THE SORT OF PERSON WHO DOESN'T CARE ABOUT WHAT PEOPLE THINK OF YOU	0.024643
S2AQ10	HOW OFTEN DRANK ENOUGH TO FEEL INTOXICATED IN LAST 12 MONTHS	0.024418
S10Q1A16	THE KIND OF PERSON WHO FOCUSES ON DETAILS/ORDER/ORGANIZATION OR LIKES TO MAKE LISTS AND SCHEDULES	0.023065
CHLD0_17	NUMBER OF CHILDREN UNDER AGE 18 IN HOUSEHOLD	0.022352
S10Q1A20	OTHERS THINK YOU HAVE UNREASONABLY HIGH STANDARDS/MORALS/IDEAS ABOUT RIGHT AND WRONG	0.022323
S10Q1A43	ARE THERE VERY FEW PEOPLE YOU'RE REALLY CLOSE TO OUTSIDE OF IMMEDIATE FAMILY	0.022088
S10Q1A25	HAVE OTHERS TOLD YOU THAT YOU ARE STUBBORN OR RIGID	0.021601
S11AQ1A2	EVER STAY OUT LATE AT NIGHT EVEN THOUGH PARENTS TOLD YOU TO STAY HOME	0.020961
NUMPER18	NUMBER OF PERSONS 18 YEARS AND OLDER IN HOUSEHOLD	0.020405
S11AQ1A1	OFTEN CUT CLASS, NOT GO TO CLASS OR GO TO SCHOOL AND LEAVE WITHOUT PERMISSION	0.020345
S10Q1A45	WOULD BE JUST HAPPY WITHOUT HAVING ANY CLOSE RELATIONSHIP	0.020180
NUMREL18	NUMBER OF RELATED PERSONS 18 YEARS AND OLDER IN HOUSEHOLD	0.019861
S2AQ5G	HOW OFTEN DRANK 5+ BEERS IN LAST 12 MONTHS	0.019397
S3AQ3B1	USUAL FREQUENCY WHEN SMOKED CIGARETTES	0.019338
S11BQ1	BLOOD/NATURAL FATHER EVER HAD BEHAVIOR PROBLEMS	0.018869
S10Q1A47	HAVE ALMOST ALWAYS PREFERRED TO DO THINGS ALONE RATHER THAN WITH OTHERS	0.018636
S11AQ1A25	EVER DO SOMETHING YOU COULD HAVE BEEN ARRESTED FOR, REGARDLESS OF WHETHER YOU WERE CAUGHT OR NOT	0.018017
S10Q1A22	HARD TO LET OTHERS HELP IF THEY DON'T AGREE TO DO THINGS EXACTLY THE WAY YOU WANT	0.017820
DGSTATUS	DRUG USE STATUS	0.017671
S1Q1D5	"WHITE" CHECKED IN MULTIRACE CODE	0.017200
S6Q1	HAD PANIC ATTACK, SUDDENLY FELT FRIGHTENED/OVERWHELMED/NERVOUS AS IF IN GREAT DANGER BUT WERE NOT	0.016674
S11AQ1A14	EVER DO THINGS THAT COULD EASILY HAVE HURT YOU OR SOMEONE ELSE, LIKE SPEEDING OR DRIVING AFTER HAVING TOO MUCH TO DRINK	0.016527
S2AQ12F	HOW OFTEN DROVE MOTOR VEHICLE AFTER 3+ DRINKS IN LAST 12 MONTHS	0.016235
S10Q1A58	FLIRT A LOT	0.015546
S3AQ3B2	USUAL FREQUENCY WHEN SMOKED CIGARS	0.014674
S2AQ9	HOW OFTEN DRANK 4+ DRINKS OF ANY ALCOHOL IN LAST 12 MONTHS (WOMEN ONLY)	0.014626
S9Q1A	EVER HAD 6+ MONTH PERIOD FELT TENSE/NERVOUS/WORRIED MOST OF TIME	0.014188
S11AQ1A22	EVER SHOPLIFT	0.014131
S10Q1A3	FIND IT HARD TO BE "OPEN" EVEN WITH PEOPLE YOU ARE CLOSE TO	0.013714
S10Q1A32	OFTEN GET ANGRY OR LASH OUT WHEN SOMEONE CRITICIZES OR INSULTS YOU	0.013670
S10Q1A46	TAKE LITTLE PLEASURE IN BEING WITH OTHERS	0.012579
S11AQ1A15	EVER GET MORE THAN 3 TICKETS FOR RECKLESS/CARELESS DRIVING, SPEEDING, OR CAUSING AN ACCIDENT	0.012403

The code and output is included below

Confusion Matrix: 

[[11983   892]

 [ 3540   823]]

Accuracy Score is:  0.742893607147

The quick and dirty on getting a Hadoop cluster up and running

The last time I tested out a hadoop cluster, it was about four years ago. At the time, the setup was manual and I used two machines to set up a two node cluster. I was able to run map reduce jobs and test the system out. Fast forward to 2017 - there are a lot more animals in the Hadoop circus and cloudera has made it very convenient by having various options to test - a quickstart VM, docker images etc. Still getting the system up and running and executing the first sqoop job took some effort. I have documented what I did including the tweaks so anyone running into the same can get help. Here are the steps:

1. Get a system ready - at least 16GB or RAM. I had a Linux Mint box on hand that I used. Mint is generally similar to Ubuntu but you do have to watch out for version specific instructions. 
2. The next step is to install Docker. For Linux Mint 18, the steps here from Simon Hardy came really handy.
3. Cloudera provides a Docker Quickstart container - do not use that. There is a lot of documentation and links on that and its quite easy to go down that path. A better option is to use the Cloudera Clusterdock. Clusterdock is a multi-node cluster deployment on the same Docker host (by default it does a two node). The clusterdock documentation is very useful but there are a few catches that I will note here. There are few other links on clusterdock here:
1. The cloudera online tutorial is based on the quickstart docker container or the quickstart VM. There are several dependencies including a mysql database, flume files that are used in the demo etc. I would suggest that you keep that container also around for a bit and copy the data as needed to the clusterdock nodes. The clusterdock setup is much more stable in the cloudera manager and the slight inconvenience (occasional hardware freezing) may be worth it. After launching the quickstart container, you may simply tar the /opt/examples folder and the mysql retail_db database and transfer that to the host machine using the docker cp command and then kill the quickstart container.
2. The clusterdock.sh script on the cloudera website lacks 'sudo' in a couple of places. Be aware of that. Its easy to spot that in case it causes a problem. For example the ssh function has this problem.
3. I wanted to run the sqoop command and for that I needed a myql database to connect to. There was one I installed on the host machine. It was almost impossible or looked very time consuming to connect the clusterdock containers to the host mysql. You run into a docker networking issue. The easy way out is to install a mysql docker container and put it on the same user defined network that the clusterdock nodes use. Note that you have to force the IP and the network on the mysql container to do that and also map the mysql port 3306 so its open for access. Else you will waste a lot of time!
4. Next step is to ssh on the master node and run a sqoop job. At this time you will run into a lot of permissions issues if you are not careful with where you are storing the target imported files. sqoop will generally report exceptions stack trace with these permissions errors. Best is to google and fix the any paths you give to the sqoop command.
5. You also need to copy the mysql jar file in the sqoop lib folders. Easiest way is to get it on the host box and then use the docker cp commands to move it to the desired location.
6. That should do it - get you past the sqoop step and then you can run a query in Hue.

Analyzing Gapminder Data

Founded in Stockholm by Ola Rosling, Anna Rosling Rönnlund and Hans Rosling, GapMinder is a non-profit venture promoting sustainable global development and achievement of the United Nations Millennium Development Goals. It seeks to increase the use and understanding of statistics about social, economic, and environmental development at local, national, and global levels. Since its conception in 2005, Gapminder has grown to include over 200 indicators, including gross domestic product, total employment rate, and estimated HIV prevalence. Gapminder contains data for all 192 UN members, aggregating data for Serbia and Montenegro. Additionally, it includes data for 24 other areas, generating a total of 215 areas.

GapMinder collects data from a handful of sources, including the Institute for Health Metrics and Evaulation, US Census Bureau’s International Database, United Nations Statistics Division, and the World Bank.

Decision tree analysis was performed (using python) to test nonlinear relationships among a series of explanatory variables (0-13 below) and a binary, categorical response variable (above average suicides per 100,000 in 2005). Note that python does not support pruning. The code is included below the analysis.

Index	Variable	Description
X[0]	income per person	2010 Gross Domestic Product per capita in constant 2000 US$
X[1]	alcohol consumption	2008 alcohol consumption per adult (age 15+),
X[2]	armed forces rate	Armed forces personnel (% of total labor force)
X[3]	breast cancer per 100th	2002 breast cancer new cases per 100,000 female
X[4]	co2 emissions	2006 cumulative CO2 emission (metric tons)
X[5]	female employment rate	2007 female employees age 15+ (% of population)
X[6]	HIV rate	2009 estimated HIV Prevalence % - (Ages 15-49)
X[7]	internet use rate	2010 Internet users (per 100 people)
X[8]	oil per person	2010 oil Consumption per capita (tonnes per year and person)
X[9]	polity score	2009 Democracy score (Polity)
X[10]	residential electricity consumption per person	2008 residential electricity consumption, per person (kWh)
X[11]	employment rate	2007 total employees age 15+ (% of population)
X[12]	urban rate	2008 urban population (% of total)
response	suicide per 100th	Is the 2005 Suicide, age adjusted, per 100 000above average? (True or False)

Alcohol consumption was the first variable to separate the sample into two subgroups. The threshold was identified to be 16.19 litres and countries where it was higher than that had a likelihood of above average suicide rate. Next one was the amount of electricity consumed in residences. If the consumption was below 230 kwh, there was a likelihood of higher suicide rates. This was followed by armed forces rates - if less than 41% of the population was part of armed forces, the likelihood was higher. Alcohol consumption and employment rate comes next - higher amounts of alcohol consumption (higher than 14%) and a low employment rate (lower than 49%) had the likelihood of an above average suicide rate. Countries with low alcohol consumption (14%) and low female employment rates (less than 56%) had lower likelihood of a below average suicide rate. However, if this is combined with low internet use (lower than 85%) this led to likelihood of above average suicide rates.

The total model classified 64% of the sample correctly.